1 /*
2  * pseries CPU Hotplug infrastructure.
3  *
4  * Split out from arch/powerpc/platforms/pseries/setup.c
5  *  arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
6  *
7  * Peter Bergner, IBM	March 2001.
8  * Copyright (C) 2001 IBM.
9  * Dave Engebretsen, Peter Bergner, and
10  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
11  * Plus various changes from other IBM teams...
12  *
13  * Copyright (C) 2006 Michael Ellerman, IBM Corporation
14  *
15  *      This program is free software; you can redistribute it and/or
16  *      modify it under the terms of the GNU General Public License
17  *      as published by the Free Software Foundation; either version
18  *      2 of the License, or (at your option) any later version.
19  */
20 
21 #include <linux/kernel.h>
22 #include <linux/interrupt.h>
23 #include <linux/delay.h>
24 #include <linux/sched.h>	/* for idle_task_exit */
25 #include <linux/cpu.h>
26 #include <linux/of.h>
27 #include <asm/prom.h>
28 #include <asm/rtas.h>
29 #include <asm/firmware.h>
30 #include <asm/machdep.h>
31 #include <asm/vdso_datapage.h>
32 #include <asm/xics.h>
33 #include "plpar_wrappers.h"
34 #include "offline_states.h"
35 
36 /* This version can't take the spinlock, because it never returns */
37 static struct rtas_args rtas_stop_self_args = {
38 	.token = RTAS_UNKNOWN_SERVICE,
39 	.nargs = 0,
40 	.nret = 1,
41 	.rets = &rtas_stop_self_args.args[0],
42 };
43 
44 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
45 							CPU_STATE_OFFLINE;
46 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
47 
48 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
49 
50 static int cede_offline_enabled __read_mostly = 1;
51 
52 /*
53  * Enable/disable cede_offline when available.
54  */
55 static int __init setup_cede_offline(char *str)
56 {
57 	if (!strcmp(str, "off"))
58 		cede_offline_enabled = 0;
59 	else if (!strcmp(str, "on"))
60 		cede_offline_enabled = 1;
61 	else
62 		return 0;
63 	return 1;
64 }
65 
66 __setup("cede_offline=", setup_cede_offline);
67 
68 enum cpu_state_vals get_cpu_current_state(int cpu)
69 {
70 	return per_cpu(current_state, cpu);
71 }
72 
73 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
74 {
75 	per_cpu(current_state, cpu) = state;
76 }
77 
78 enum cpu_state_vals get_preferred_offline_state(int cpu)
79 {
80 	return per_cpu(preferred_offline_state, cpu);
81 }
82 
83 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
84 {
85 	per_cpu(preferred_offline_state, cpu) = state;
86 }
87 
88 void set_default_offline_state(int cpu)
89 {
90 	per_cpu(preferred_offline_state, cpu) = default_offline_state;
91 }
92 
93 static void rtas_stop_self(void)
94 {
95 	struct rtas_args *args = &rtas_stop_self_args;
96 
97 	local_irq_disable();
98 
99 	BUG_ON(args->token == RTAS_UNKNOWN_SERVICE);
100 
101 	printk("cpu %u (hwid %u) Ready to die...\n",
102 	       smp_processor_id(), hard_smp_processor_id());
103 	enter_rtas(__pa(args));
104 
105 	panic("Alas, I survived.\n");
106 }
107 
108 static void pseries_mach_cpu_die(void)
109 {
110 	unsigned int cpu = smp_processor_id();
111 	unsigned int hwcpu = hard_smp_processor_id();
112 	u8 cede_latency_hint = 0;
113 
114 	local_irq_disable();
115 	idle_task_exit();
116 	xics_teardown_cpu();
117 
118 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
119 		set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
120 		if (ppc_md.suspend_disable_cpu)
121 			ppc_md.suspend_disable_cpu();
122 
123 		cede_latency_hint = 2;
124 
125 		get_lppaca()->idle = 1;
126 		if (!get_lppaca()->shared_proc)
127 			get_lppaca()->donate_dedicated_cpu = 1;
128 
129 		while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
130 			extended_cede_processor(cede_latency_hint);
131 		}
132 
133 		if (!get_lppaca()->shared_proc)
134 			get_lppaca()->donate_dedicated_cpu = 0;
135 		get_lppaca()->idle = 0;
136 
137 		if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
138 			unregister_slb_shadow(hwcpu);
139 
140 			/*
141 			 * Call to start_secondary_resume() will not return.
142 			 * Kernel stack will be reset and start_secondary()
143 			 * will be called to continue the online operation.
144 			 */
145 			start_secondary_resume();
146 		}
147 	}
148 
149 	/* Requested state is CPU_STATE_OFFLINE at this point */
150 	WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
151 
152 	set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
153 	unregister_slb_shadow(hwcpu);
154 	rtas_stop_self();
155 
156 	/* Should never get here... */
157 	BUG();
158 	for(;;);
159 }
160 
161 static int pseries_cpu_disable(void)
162 {
163 	int cpu = smp_processor_id();
164 
165 	set_cpu_online(cpu, false);
166 	vdso_data->processorCount--;
167 
168 	/*fix boot_cpuid here*/
169 	if (cpu == boot_cpuid)
170 		boot_cpuid = cpumask_any(cpu_online_mask);
171 
172 	/* FIXME: abstract this to not be platform specific later on */
173 	xics_migrate_irqs_away();
174 	return 0;
175 }
176 
177 /*
178  * pseries_cpu_die: Wait for the cpu to die.
179  * @cpu: logical processor id of the CPU whose death we're awaiting.
180  *
181  * This function is called from the context of the thread which is performing
182  * the cpu-offline. Here we wait for long enough to allow the cpu in question
183  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
184  * notifications.
185  *
186  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
187  * self-destruct.
188  */
189 static void pseries_cpu_die(unsigned int cpu)
190 {
191 	int tries;
192 	int cpu_status = 1;
193 	unsigned int pcpu = get_hard_smp_processor_id(cpu);
194 
195 	if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
196 		cpu_status = 1;
197 		for (tries = 0; tries < 5000; tries++) {
198 			if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
199 				cpu_status = 0;
200 				break;
201 			}
202 			msleep(1);
203 		}
204 	} else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
205 
206 		for (tries = 0; tries < 25; tries++) {
207 			cpu_status = smp_query_cpu_stopped(pcpu);
208 			if (cpu_status == QCSS_STOPPED ||
209 			    cpu_status == QCSS_HARDWARE_ERROR)
210 				break;
211 			cpu_relax();
212 		}
213 	}
214 
215 	if (cpu_status != 0) {
216 		printk("Querying DEAD? cpu %i (%i) shows %i\n",
217 		       cpu, pcpu, cpu_status);
218 	}
219 
220 	/* Isolation and deallocation are definitely done by
221 	 * drslot_chrp_cpu.  If they were not they would be
222 	 * done here.  Change isolate state to Isolate and
223 	 * change allocation-state to Unusable.
224 	 */
225 	paca[cpu].cpu_start = 0;
226 }
227 
228 /*
229  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
230  * here is that a cpu device node may represent up to two logical cpus
231  * in the SMT case.  We must honor the assumption in other code that
232  * the logical ids for sibling SMT threads x and y are adjacent, such
233  * that x^1 == y and y^1 == x.
234  */
235 static int pseries_add_processor(struct device_node *np)
236 {
237 	unsigned int cpu;
238 	cpumask_var_t candidate_mask, tmp;
239 	int err = -ENOSPC, len, nthreads, i;
240 	const u32 *intserv;
241 
242 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
243 	if (!intserv)
244 		return 0;
245 
246 	zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
247 	zalloc_cpumask_var(&tmp, GFP_KERNEL);
248 
249 	nthreads = len / sizeof(u32);
250 	for (i = 0; i < nthreads; i++)
251 		cpumask_set_cpu(i, tmp);
252 
253 	cpu_maps_update_begin();
254 
255 	BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
256 
257 	/* Get a bitmap of unoccupied slots. */
258 	cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
259 	if (cpumask_empty(candidate_mask)) {
260 		/* If we get here, it most likely means that NR_CPUS is
261 		 * less than the partition's max processors setting.
262 		 */
263 		printk(KERN_ERR "Cannot add cpu %s; this system configuration"
264 		       " supports %d logical cpus.\n", np->full_name,
265 		       cpumask_weight(cpu_possible_mask));
266 		goto out_unlock;
267 	}
268 
269 	while (!cpumask_empty(tmp))
270 		if (cpumask_subset(tmp, candidate_mask))
271 			/* Found a range where we can insert the new cpu(s) */
272 			break;
273 		else
274 			cpumask_shift_left(tmp, tmp, nthreads);
275 
276 	if (cpumask_empty(tmp)) {
277 		printk(KERN_ERR "Unable to find space in cpu_present_mask for"
278 		       " processor %s with %d thread(s)\n", np->name,
279 		       nthreads);
280 		goto out_unlock;
281 	}
282 
283 	for_each_cpu(cpu, tmp) {
284 		BUG_ON(cpu_present(cpu));
285 		set_cpu_present(cpu, true);
286 		set_hard_smp_processor_id(cpu, *intserv++);
287 	}
288 	err = 0;
289 out_unlock:
290 	cpu_maps_update_done();
291 	free_cpumask_var(candidate_mask);
292 	free_cpumask_var(tmp);
293 	return err;
294 }
295 
296 /*
297  * Update the present map for a cpu node which is going away, and set
298  * the hard id in the paca(s) to -1 to be consistent with boot time
299  * convention for non-present cpus.
300  */
301 static void pseries_remove_processor(struct device_node *np)
302 {
303 	unsigned int cpu;
304 	int len, nthreads, i;
305 	const u32 *intserv;
306 
307 	intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
308 	if (!intserv)
309 		return;
310 
311 	nthreads = len / sizeof(u32);
312 
313 	cpu_maps_update_begin();
314 	for (i = 0; i < nthreads; i++) {
315 		for_each_present_cpu(cpu) {
316 			if (get_hard_smp_processor_id(cpu) != intserv[i])
317 				continue;
318 			BUG_ON(cpu_online(cpu));
319 			set_cpu_present(cpu, false);
320 			set_hard_smp_processor_id(cpu, -1);
321 			break;
322 		}
323 		if (cpu >= nr_cpu_ids)
324 			printk(KERN_WARNING "Could not find cpu to remove "
325 			       "with physical id 0x%x\n", intserv[i]);
326 	}
327 	cpu_maps_update_done();
328 }
329 
330 static int pseries_smp_notifier(struct notifier_block *nb,
331 				unsigned long action, void *node)
332 {
333 	int err = 0;
334 
335 	switch (action) {
336 	case OF_RECONFIG_ATTACH_NODE:
337 		err = pseries_add_processor(node);
338 		break;
339 	case OF_RECONFIG_DETACH_NODE:
340 		pseries_remove_processor(node);
341 		break;
342 	}
343 	return notifier_from_errno(err);
344 }
345 
346 static struct notifier_block pseries_smp_nb = {
347 	.notifier_call = pseries_smp_notifier,
348 };
349 
350 #define MAX_CEDE_LATENCY_LEVELS		4
351 #define	CEDE_LATENCY_PARAM_LENGTH	10
352 #define CEDE_LATENCY_PARAM_MAX_LENGTH	\
353 	(MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
354 #define CEDE_LATENCY_TOKEN		45
355 
356 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
357 
358 static int parse_cede_parameters(void)
359 {
360 	memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
361 	return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
362 			 NULL,
363 			 CEDE_LATENCY_TOKEN,
364 			 __pa(cede_parameters),
365 			 CEDE_LATENCY_PARAM_MAX_LENGTH);
366 }
367 
368 static int __init pseries_cpu_hotplug_init(void)
369 {
370 	struct device_node *np;
371 	const char *typep;
372 	int cpu;
373 	int qcss_tok;
374 
375 	for_each_node_by_name(np, "interrupt-controller") {
376 		typep = of_get_property(np, "compatible", NULL);
377 		if (strstr(typep, "open-pic")) {
378 			of_node_put(np);
379 
380 			printk(KERN_INFO "CPU Hotplug not supported on "
381 				"systems using MPIC\n");
382 			return 0;
383 		}
384 	}
385 
386 	rtas_stop_self_args.token = rtas_token("stop-self");
387 	qcss_tok = rtas_token("query-cpu-stopped-state");
388 
389 	if (rtas_stop_self_args.token == RTAS_UNKNOWN_SERVICE ||
390 			qcss_tok == RTAS_UNKNOWN_SERVICE) {
391 		printk(KERN_INFO "CPU Hotplug not supported by firmware "
392 				"- disabling.\n");
393 		return 0;
394 	}
395 
396 	ppc_md.cpu_die = pseries_mach_cpu_die;
397 	smp_ops->cpu_disable = pseries_cpu_disable;
398 	smp_ops->cpu_die = pseries_cpu_die;
399 
400 	/* Processors can be added/removed only on LPAR */
401 	if (firmware_has_feature(FW_FEATURE_LPAR)) {
402 		of_reconfig_notifier_register(&pseries_smp_nb);
403 		cpu_maps_update_begin();
404 		if (cede_offline_enabled && parse_cede_parameters() == 0) {
405 			default_offline_state = CPU_STATE_INACTIVE;
406 			for_each_online_cpu(cpu)
407 				set_default_offline_state(cpu);
408 		}
409 		cpu_maps_update_done();
410 	}
411 
412 	return 0;
413 }
414 arch_initcall(pseries_cpu_hotplug_init);
415